Sambatra Andrianomena, Mika Rafieferantsoa, Holifidy Rapanoel, Rondro Barimalala

Tropical cyclones (TC) are one of the most destructive natural events claiming a lot of human lives and devastating coastal areas. Despite the advanced understanding of the formation of TC, prediction capabilities on the rapid intensification (RI) of TCs remain unsatisfactory. In this study, a deep learning framework using satellite images is used for the first time to identify RI events. We resort to the predictive power of VGG-like, ResNet-like and Xception architectures. The results show that the models are well capable of differentiating RI from non-RI events (roc-auc> 0.86), with a probability of detection (POD) > 0.83and high fractional improvement over a random guess (HSS) > 0.57. The False Alarm Rate (FAR) is less than 0.23 on average.By considering only the best performance of the learners,roc-aucis maximized to 0.878 for VGG , 0.874 for ResNet and 0.911for Xception; FAR decreases to 0.218 for VGG, 0.209 for ResNet and 0.182 for Xception, and POD are 0.864, 0.835 and 0.888for the three models respectively. The trained models can be deployed in a real world scenario to help mitigate the further risks engendered by a TC going through a phase of rapid intensification.